Convergence to Steady State in LLM-Generated Ontological Concepts

Large Language Models (LLMs) offer a scalable method to generate candidate concepts for ontology expansion. However, the outputs depend on LLMs' sampling parameters such as "temperature." A high temperature leads to increased hallucinations. We consider two hypotheses: (1) LLM outputs for higher temperatures should be supersets of corresponding outputs for lower temperatures; (2) Repeated LLM outputs for lower temperatures should converge to a steady state faster than for higher temperatures. A steady state is defined as k consecutive iterations with no new concepts generated (k-convergence). In this study, we analyzed Environmental Determinants of Health (EnDOH) concepts retrieved at sampling temperatures ranging from 0.1 to 1.0 using concept-structured prompting. Contrary to expectations, higher temperatures do not produce strict supersets of lower temperature outputs. To assess convergence towards a steady state we iteratively executed fixed prompts at different temperatures. Our results confirm that lower temperatures indeed lead to faster convergence.